Data assimilation into numerical models should be both computationally fast and physically meaningful, in order to be applicable in online environmental surveillance. We present a way to improve assimilation for computationally intensive models, based on non-stationary kriging and a separable space–time covariance function. The method is illustrated with significant wave height data. The covariance function is expressed as a collection of fields: each one is obtained as the empirical covariance between the studied property(significant wave height in log-scale)at a pixel where a measurement is located (a wave-buoy is available)and the same parameter at every other pixel of thef ield. These covariances are computed from the available history of forecasts. The method provides a set of weights, that can be mapped for each measuring location, and that do not vary with time. Resulting weights may be used in a weighted average of the differences between the forecast and measured parameter. In the case presented, these weights may show long-range connection patterns, such as between the Catalan coast and the eastern coast of Sardinia, associated to common prevailing meteo-oceanographic conditions. When such patterns are considered as non-informative of the present situation, it is always possible to diminish their influence by relaxing the covariance maps.